Communications in Statistics - Theory and Methods

ISSN: 0361-0926 (Print) 1532-415X (Online) Journal homepage: https://www.tandfonline.com/loi/lsta20

Modeling failure time data by lehman alternatives

Ramesh C. Gupta , Pushpa L Gupta & Rameshwar D. Gupta

To cite this article: Ramesh C. Gupta , Pushpa L Gupta & Rameshwar D. Gupta (1998) Modeling
failure time data by lehman alternatives, Communications in Statistics - Theory and Methods, 27:4,
887-904, DOI: 10.1080/03610929808832134

To link to this article: https://doi.org/10.1080/03610929808832134

Published online: 27 Jun 2007.

Submit your article to this journal

Article views: 788

View related articles

Citing articles: 115 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=lsta20
COMMUN. STATIST.-THEORY METH., 27(4), 887-904 (1998)

MODELING FAILURE TIME DATA

BY LEHMAN ALTERNATIVES

Ramesh C . Gupta Rarneshwar D. Gupta

and
Pushpa L. Gupta

Department of Mathematics Department of Math; Stat.

and Statistics and Computer Science
University of Maine University of New Brunswick
Orono, ME 04469-5752 Saint John, N.B., Canada E2L4L5

Keywords and phrases: Monotonic and non-monotonic failure rates,

exponentiated Weibull, exponentiated Pareto, exponentiated gamma, order
relations

ABSTRACT

The proportional hazards model has been extensively used in the literature
to model failure time data. In this paper we propose to model failure time data by
F*(t) = [ ~ ( t ) ] where
' F(t) is the baseline distribution function and 8 is a

positive real number. This model gives rise to monotonic as well as non-
monotonic failure rates even though the baseline failure rate is monotonic. The
monotonicity of the failure rates are studied, in general, and some order relations
are examined. Some examples including exponentiated Weibull, exponential,
gamma and Pareto distributions are investigated in detail.

887

Copyright O 1998 by Marcel Dekker, Inc.

888 G"UPTA, GUPTA, AND GUPTA

1~ INTRODUCTION

Cox's (1972) proportional hazard model (PHM) for survival analysis has
gained widespread attention from both applied and theoretical statisticians to
model failure time data. For example in the comparison of two survival functions

PI and F2 in a clinical trial, the PHM assumes that FI(t) = [F2(t)]6, for some
constant e. In this setting the parameter e has the interpretation of relative risk
and has intuitive appeal as a descriptive measure in survival. The model also
implies~ that the two hazard rates corresponding to the distribution functions
F 1 and F 2 are proportional and thus have the same monotonic properties.

In this paper we propose to model failure time data by

F*(t) = [F(t)]6, where F(t) is the baseline distribution function and e is a

positive real number and study its properties. This model has also been called in 9

the literature, as Lehman alternatives, when e is in fact a positive integer.

Lehman (1953) has studied such alternatives to define various nonparametric
hypotheses and has computed the approximate power of certain rank tests using
large sample theory.
The proposed model is flexible enough to accommodate both monotonic
as well as non-monotonic failure rates even though the baseline failure rate is
monotonic. Modeling survival data by non-monotonic failure rates is desirable,
for example, when the course of the disease is such that mortality reaches a peak
after some finite period and then slowly declines. Recently Mudholkar and
Srivastava (1993), Mudholkar et.a!. (1995) and Mudholkar and Hutson (1996)
have modeled various failure time data sets with the proposed model with Weibull
as the baseline distribution. In two of these papers, the authors have examined the
maximum likelihood estimation of the so~called exponentiated Weibull family
and studied the goodness of fit of the null hypothesis e= 1 within the

exponentiated-Weibull model. In Mudholkar et.a!' (1995), the usefulness and

flexibility of the family is illustrated by reanalyzing five classical data sets on bus
MODELING FAILURE TIME DATA 889

motor failures from Davis (1952) that are typical of data in repair-reuse situations
and Efron's (1988) data pertaining to a head and neck cancer clinical trial.
In the present paper, in section 2 the monotonicity of failure rates are
studied, in general, in relation to the monotonicity of the baseline hazard. A set of
suffi~ient conditions are provided for F* to be IFR (DFR), increasing failure rate
(decreasing failure rate) when F is IFR (DFR). In section 3, we study some order
relations between the failure rates and mean residual life functions (MRLF).
Section 4 contains the detailed analysis for the cases when the baseline model is
exponential, Weibull, gamma and Pareto. In each case, conditions are obtained on
the parameters for F* to be IFR, DFR, bath tub or upside bath tub (to be defined
later).

2. THE PROPOSED MODEL AND ITS MONOTONIC PROPERTIES

Let T be a non-negative random variable denoting the life length of a component

having distribution function F(t) with F(O) = 0 and the pdfj(t). Then the failure
rate of T is given by

ret) = f(t)IF(t),

where F(t) =1 - F(t) is the survival function of T.

Let T* be a non-negative random variable such that its distribution function F*(t)
is an exponentiated function of F i.e.

(2.1)

Therefore, the pdf of T* is

f*(t) = eFe-1(t)f(t) t > 0, e > o. (2.2)

The failure rate of T* is given by

r*(t) = f*(t)IF*(t) = e ret) get), (2.3)

where
Fe-I(t) - Fe(t)
get) (2.4)
1 -) Fe(t)

In order to examine the monotonic properties of r*(t), we look at

 GUPTA, GUPTA, AND GUPTA

It can be verified that g(t) is an increasing (decreasing) function of

r for 8 > I(< I). Thus we have the following result.
Theorem 2.1
(a) If 8 > 1 and F E IFR (increasing failure rate distribution class), then

(b) If 8 < 1 and F E DFR (decreasing failure rate distribution class), then
F* E DFR.
Remark. The condition 8 > (<) 1 is sufficient and not necessary for
F *EIFR(DFR) if F E IFR(DFR).
In order to obtain more general conditions, we define the
following classes of failure rates for a general r(t), see Glaser (1980).
1. If rl(t) > 0 for all t, we say that F E IFR.
2. If rl(t) < 0 for all t, we say that F E DFR.
3. If there exists a t * > 0 such that rl(t) < 0 for all t E (O,t*),
rl(t*) = 0 and rl(t*) > 0 for all t > t*, we say that F has a bath tub
*
shaped failure rate and F E B (the class of bath tub shaped failure rate).
4. If there exists a t * > 0 such that rl(t) > 0 for all
t E. (0, t *), r'(t *) = 0 and r'(t *) < 0 for all t > t *, we say that F has
an upside bath tub shaped failure rate and F E U (the class of upside bath
tub shaped failure rate).
To determine the nature of the failure rate, we proceed as follows:
Define
rl(t) = -f(t)lf(t),
whereflt) is the pdf corresponding to r(t). The following result is due to Glaser
(1980).
MODELING FAILURE TIME DATA

Lemma 2.1
If ql(t) > 0 for all t > 0,then F E IFR.
If ql(t) < 0 for all t > 0, then F E DFR.

Suppose there exists to > 0 such that

ql(t)<O for all t E (O,to),qt(to)= O and q' (t) > 0 for all t > to and

(i) if limit f(t) = m, then F E B.

1-0.

(ii) if limit f(t) = 0 then F E IFR.

1-0.

Suppose there exists a to > 0 such that

ql(t) > 0 for all t E (0, to), q1(t0) = 0 and

qf(t) < 0 for all t > to and

(i) if limit At) = 0, then F E U.

1-0'

(ii) if limit At) = -, then F E DFR.

1-0'

We now apply the above Lemma to F*(t) = ~ ' ( t ) .

We have
f+'(t) = 8(8 - 1) ~ ' - ' ( t ) f2(t) + 8 ~ ' - ' ( t )f(t).
This gives

d2 1 d2
Thusif 8 > I and In F(t) < - - Inf(t)forall t > 0 , then
dt (1 - 8 ) d t 2

F * E IFR.

d2
Also if 8 < 1 and -In F(t)>-
1 d2
-lnfft) for all t > 0 , then
dt (1 -8) d t 2
892 GUPTA, GUPTA, AND GUPTA

F * E IFR. Similar conditions can be obtained for F* to belong to DFR.

3. SOME ORDER RELATIONS

In this section, we shall study some order relations between the

distribution F and F*.
1. Likelihood Ratio Ordering
Let X and Y denote continuous non-negative random variables with
respective densities f(.) and g(.). We say that X 2 Y if f(t)lg(t) is a
U1

nondecreasing function oft.

2. Failure Rate Ordering

We say that X < Y if rJt) s rAt) for all t > 0 .
FR

Since X 2 Y implies rJt) r; r&t) for all t, we conclude that, if

LR

8 > 1 , r*(t)< r(t) forall t > Oandif 8 < 1, r * ( t ) z r(t) forall t > 0 .
This can also be seen by considering X* as a weighted version of X with
weight function w(t) = 8 ~ ' - ' ( t ) . Since 6 > 1 (< 1) implies that w(t) is
increasing (decreasing) function o f t , we have
r*(t) 5 r(t) if 8 > 1 and r*(t) z r(t) if 0 < 1,
see Gupta and Kirmani (1990) for details.
3. Mean Residual Life Ordering
The mean residual life (MRL) or life expectancy of a random variable X is
defined as
MODELING FAILURE TIME DATA

p(t) = E(X - tjX > t).

The failure rate and the mean residual life function are connected by the
relation

We say that X s Y if pJt) 5 p At) for all t > 0. Since failure rate ordering
MRL

implies MRL ordering, see Gupta and Kirmani (1987), we have

p *(t) 2 ~ ( t for
) all t if 0> 1
and p *(t) 1; p(t) for all t if 0 < 1.
4. Diswrsive Ordering
For any two distributions F and G, F is said to be dispersed with respect to
G (F < G), if the following equivalent conditions are satisfied.
(a) F-I(P) - F -'(a) 2 G "(P) - G -'(a) whenever 0 < a < p c 1.

(b) G -'F(x) - x is nondecreasing in x, see Shaked (1982).

(c) g(G -l(u)) s f(F -l(u)) whenever u E. (0, 1) and f and g are the
densities corresponding to F and G.

(d) x - F "G(x) is nondecreasing in x.

We now state a relationship between the failure rate ordering and the dispersive
ordering.
disp
If rdt) 2 rG(t) for all t 2 0 and F o r G is DFR, then < G ; see
Bartoszewicz(l987).
disp
Thus in our case if 0 < 1 and F E DFR, then * < F.

5. Order Statistics Ordering

disp
For any two distributions F and G, suppose < G . Also suppose
Xi,,, i = 1, 2...n and Y,,", i = 1, 2...n are the order statistics for F and G

respectively. Then
894 GUPTA, GUPTA, AND GUPTA

In particular
COY (YjXn,Y,.,,) 2 cov(Xj:,, Xi:,) and var(YJ z var(Xi:,).

d) For all convex functions 4,

E [ 4 ( Y - E(Y))I 2 E t 4 ( X - E(X))I.
e) For all convex functions 4 ,
E [ 4 ( Y j : , - E(Yin))I 2 E[@(X,:, - E(Xi:,))I, i = 1,2...n.

For details, see Bartoszewicz (1985). Thus in our case if 0 < 1 and F E DFR,

wehave F* disp
< F and all the results listed

above can be obtained.

4. SOME EXAMPLES

4.1 Exponentiated Weibull Family

Here f ( t ) = arm-'e - l a , F(t) = 1 - e - l a , r(t) = ata-'.

F*(t) = (1 - e f*(t) = 0 ( 1 - e -'a)e-' a fa-' e -la.

Thus from Theorem 2.1,

1. If 0 > 1 and a > 1, then F* E IFR.
2. I f 0 < 1 and a < 1, then F* E DFR.
If these conditions are not satisfied, we proceed as follows:

limit f * ( r ) = Oa limit e - I a limit ta-'(1 - e

1-0. 1-0. 1-0.
MODELING FAILURE TIME DATA

= 0 a limit rue-'
r-0-

It can be verified that in this case

q ( t )= - 0 - a t
a - 2 .p
e
I - a t a + ( a - 1 )- atue-"

Analysis of q * ' ( t )leads to the following result, see Mudholkar et.al. (1995).

Theorem 4.1

The failure rate is

a) IFR if a 2 1 and a0 2 1.

b) DFR if a < 1 and a0 5 1.

c) of the type B if a > 1 and a0 < 1 and

d) of t h e t y p e u i f a < 1 and a0 > 1.

The monotonicities are strict except for the negative exponential distribution

corresponding to a = 0 = 1 .

Exponentiated-Exponential Family

In this case a = 1 and

Clearly q*'(t)> 0 for 0 > 1 and q*'(t) < 0 for 0 < 1 .

Then F* E IFRif 0>1 and F * E DFRfor 0<1
896 GUPTA, GUPTA, AND GUPTA

4.2 Exponentiated-Pareto Family

a
Here f(t) = a ( l + t)-(""), F(r)= 1 - ( 1 + t)-",r(t)=-,a>0
I +t
Here F E DFR.

F*(t) = [ I - ( I + t ) - a ] e , a , O , t > ~

f*(t) = 8a[1 - ( I + t)-"1'-'(1+ t)-('+').

It can be verified that in this case

q *'(t)<Ofor all t>O if 8 < 1 i.e. if 8 .s 1 , F* E DFR.
For the case 0 > 1, we first find the solutions of q * ' ( t ) = 0 . Let

y = (1 +r)'. Then the equation q * ' ( t )= 0 can be reduced to

( ~ 1 + l ) ~ ~ - ( a + 1 ) ( 2 + a 0 - a ) ~ + ( aThesolutionsare
8+1)=0.

When 0 > I , one of the root is positive. Also, since the product of the roots is
positive, the other root is also positive. Therefore, q * ' ( t )= 0 at

a 8 [ ( l +t12 - l l e - '
Now f*(t) =
(I + t)"u8

Since 0> 1 and a>O,limitf*(t)=O.

1-0.

We now show that q * ' ( t ) = O has only one positive root. Suppose both the roots
in t are positive. Since the root corresponding to the positive sign is positive, the
root corresponding to the negative sign is assumed to be positive also, i.e.
 MODELING FAILURE TIME DATA

or - 4 a > 4 - 4 ( 1 + a ) < O ,
which is not true. So if both the roots are positive, then
1 + a < 0 , a contradiction.
4.3 Exponentiated-Gamma Family
Here A t ) = te -',t 2 0, F(t) = 1 - e -'- te -'

Here F E IFR,

F * ( ~ ) = (- eI - ' - t e - ' ) e ,

f*(t)=ete"(I - e - ' - t e - ? e - ' .

1 (6-1)(er(l-t)-1)
It can be verified that in this case q *'(t) = --
t2 ( e l - ( 1 +t)12

NOW q * ' ( r ) >O iff (' - '(I - t ) - <-!- for a11 t.

(e1-(l+t))2 t2

e1(1 - t ) - I
Since i 0 for all t ,
( e ' - ( I +t)12

q * ' ( t ) > O iff 0 > 1 + ( ' - + I 2 = I - ~ ( twhere

) w(t) = ( e ' - 1 - t)'
t 2 [ e' ( 1 - t ) - 1] t 2 ( t e ' - e'+ 1 )

w ( t ) is non negative and is an increasing function oft, which can be seen as

follows:
898 GUPTA, GUPTA, AND GUPTA

We write w l ( t ) as

el-I - t
wi(t) = g(t)
(te' - e' + l)'t3

where g(t) = 2 ( t e f - e f + l ) ' - r 2 e r ( e ' - 1- t ) .

In order to show that w(t) is increasing, it is enough to show that g(t) > 0 for all

t > 0. Consider g l ( t ) = te 'h(t),where

h(t) = e 1 ( 2 t - 6 ) + t 2 + 4t + 6with

hl(t) = 2[ef(r-2)+t+2]

h n ( t ) = 2(te ' - e + 1)

hn'(t) = 2te '.

For t > 0, h"'(t) > 0 . This implies that h"(t) is increasing. Since
h"(0) = 0, h"(t) > 0 for all t > 0 . Arguing in the same way in succession, we
can show that g(t) > 0 for all t.

Since the minimum value of w(t) = 112, ~ ' ( t>) 0 if and only if
0 > 112. Hence F* E IFR iff 8 > 112.
We shall now show that in case 0 5 112, F* E B. We have

. Since w(t) is increasing and its minimum value is 112,

there exists a point to such that w(to) = I - 8 and hence q *'(to) = 0 . By

simple calculations, it can be shown that r, is a point of minima. Also it can be

verified that limit f *(r) = -. Hence F * E B if 8 5 112.

1-0.

4.4 Exponentiated Gamma Family (General Case)

Here we consider Exponentiated-Gamma ( a = k + 1, P = l ) , where k is a
positive integer. Then
MODELING FAILURE TIME DATA

F*(t) = ( ~ ( t ) )and
~ , f*(t) = 0 J t ) ( ~ ( t ) ) ,~ - '

It can be verified, in this case, that

where w ( t ) = [= ): 2
/tk+'Z
1
" (i-k)
+ l!
I. Thus 9 * ' ( t ) > 0 iff 0 > I - k ! k w ( t ) .

It can be verified that w ( t ) is non-negative and is an increasing function of

t (for proof, see the appendix). Since the minimum value of
k!k =
- k -- -
-
w(t) = -, q * ! ( t ) > o i f f ~ >I - 1 - I . Hence
(k+l)! (k+l)! k+l k+l

1
We now show that in case 0 s -, F* E B . Recall that
k+l
. Since w ( t ) is increasing and its minimum value is
I , there exists a point to such that w(tJ = I - 0 and q * ' ( t , ) = 0 . B y
( k +l ) !
simple calculations, it can be shown that to is a point of minima. Also it can be
seen that limit f * ( t ) = m. Hence F* E B if 8 < tl(k+ I )
1--0.
900 GUPTA, GUPTA, AND GUPTA

5. APPENDIX

Let w ( t ) =

We want to show that w ( t ) is an increasing function o f t .

Let

then

and t z l ( t ) + ~ ( t =)
"
-. ( i + l ) t' Then w(t) can be rewritten as
l=o (k+l+i)!

Since z(t) > 0 for all t > 0, in order to show that w(t) is increasing for t > 0, it
is enough to show that
2zt2 - z"z > 0 for all t > 0 , (5.4)
suppressing the argument for the sake of brevity. From (5.2) and (5.3) we find
that
MODELING FAILURE TIME DATA

where

To show (5.4), it is enough to show that ai > bi for all i.

Let us observe the following two results:

1. The sums of the numerators in (5.5) and (5.6) are equal, i.e.,

The above can be proved with the use of well-known formulas such as

2. If i = 0 then, from (5.5) and (5.6),

Since k + 2 < k+3, [(k+2)!12< (k+l)!(k+3)!.This implies that a, > b,.

Now let i be positive. By combining terms with the same denominators in

(5.5) and (5.6), ai and bi can be written as
902 GUPTA. GUPTA, AND GUPTA

i- l
means the integer part of -; the last term in (5.7) and in (5.8) is
2

absent if i is odd. The sums of the numerators in (5.7) and (5.8) are equal since
they coincide with the sums of the numerators in (5.5) and (5.6).
Hence

where c - , = -(i+l)(i+2),

1)-j(j+l)
cI. = 4(j+l)(i-j+l)-(i-j)(i-j+

= -6j2+6ji+4+3i-j2,0 < j < [i/2],

cn = (;+1)(i/2+2), ifiiseven

and

We observe that c., < 0 , c,,? > 0 and c, are increasing with j for 0 s j < -
[:I
as c J + -cl
, =
(; )
12 ----I > 0. Let P be the smallestjsuch that cJ z 0 , so

c -,,...,c,., are negative and cr....,clln,are non-negative. Since

(k+2+j)!(k+2+i-j)!> (k+2+P)!(k+2+i-P)!and cj < 0 for j < P,

and c, r 0 for j
(k+2+j)!(k+2+i-j)!< (k+2+Q)!(k+2+i-P)! 2 P,
MODELING FAILURE TIME DATA

we have
C
I 1 "1 for all j,
(k+2+j)(k+2+i-j)! (k+2+B)!(k+2+i-Q)!

where at least some inequalities are strict.

Hence from (5.9) we obtain, in light of the above result,

Hence, a, - b, > 0 for all i. Thus (5.4) is true and w ( t ) ,given by (5.1), is an

increasing function for t > 0 .

ACKNOWLEDGEMENTS

The work of the first two authors was supported, inpart, by a travel grant
from the Canadian American Center of the University of Maine. The work of the
third author was supported, in part, by the NSERC grant GOP0004850.
The authors are thankful to Dr. Sergey Lvin for his help in proving the result
given in the Appendix.
BIBLIOGRAPHY

Bartoszewicz, J. (1985). Moment inequalities for order statistics from ordered

families of distributions. Metrika 32. 383-389.

Cox, D.R. (1972). Regression models and life tables (with discussion). Journal of
the Royal Statistical Society, series B, 34, 187-220.

Davis, D.J. (1952). An analysis of some failure data. Journal of the American
Statistical Association 47, 1 13-150.

Efron, B. (1988). Logistic regression survival analysis and the Kaplan-Meier

curve. Journal of the American Statistical Association 83. 4 14-425.

Glaser, R.E. (1980). Bathtub and related failure rate characterizations. Journal of
the American Statistical Association 75, 667-672.

Gupta, R.C. and Kirmani, S.N.U.A. (1990). The role of weighted distribut~onsin
stochastic modeling. Communications in Statistics - Theory and methods,
19(9), 3 147-3162.
904 GUPTA, GUPTA, AND GUPTA

Gupta, R.C. and Kirmani, S.N.U.A. (1987). On order relations between reliability
measures. Stochastic Mode!s, 3(1), 149.156.

Lehman, E.L. (1953). The power of rank tests. Annals of Mathematical Statistics,
24. 28-43.

Mudholkar, G.S. and Srivastava, D.K. (1993). Exponentiated Weibull family for
analyzing bath tub failure data. IEEE Transactions on Reliability 42(2),
299-302.

Mudholkar, G.S. and Srivastava, D.K. and Freimer, M. (1995). The exponentiated
Weibull family: a reanalysis of the bus-motor-failure data. Technometrics
37(4), 436-445.

Mudholkar, G.S. and Hutson, A.D. (1996). The exponentiated Weibull family:
some properties and a flood data application. Communications in
Statistics - Theory & Methods, 25(12), 3059-3083.

Received October, 1997; Revised November, 1997.